NVIDIA's robot business is already worth $10 billion/year, and China is its best customer
NVIDIA's physical AI business is already generating roughly $10 billion a year, more than the entire humanoid robot industry combined, with Jensen Huang pointing to a tenfold path over the next decade. A meaningful share of that comes from China, which built about 90% of the world's humanoids last year. But Beijing is racing to build a domestic stack ...
Roughly $10 billion
That's the estimated annual revenue from Nvidia's physical AI business — the chips, models, and simulation software that run robots and autonomous systems — according to Wall Street Journal reporting.
CEO Jensen Huang has suggested a path to ten times that over the next decade.
For context, that $10 billion is a rounding error against Nvidia's data-center business. But it is also, on its own, larger than the combined revenue of essentially every humanoid robot company on earth. The picks-and-shovels layer is monetizing years ahead of the robots.
And a meaningful share of it is being bought by Chinese manufacturers.
The NVIDIA stack
Nvidia's physical AI offering is not one product but a full-stack play, and each layer is designed to make the next one necessary:
- GR00T — foundation models for humanoid robot control
- Cosmos — world models and synthetic data generation for training
- Isaac / Isaac Lab — simulation and reinforcement-learning environments
- Jetson — the edge compute that actually rides on the robot
The strategic logic is the same one that worked in AI training: don't sell the robot, sell the thing every robot needs. Nvidia doesn't have to pick which humanoid company wins.
It needs the category to exist. And, ideally, grow like crazy.
The China exposure
But there's a policy twist.
Washington's export restrictions have been aimed squarely at advanced AI training chips. Robotics and automotive systems have largely fallen outside that perimeter, meaning NVIDIA can still sell substantial physical AI hardware and software into China, into an ecosystem that produced roughly 90% of the world's humanoid robots last year.
Huang has been unusually direct about why that ecosystem matters. Speaking at GTC, he called China "formidable" in robotics, pointing to its strengths in "microelectronics, motors, rare earth and magnets" ... the unglamorous physical inputs that no amount of American software compensates for.
(Check out our recent story on exactly that.)
He also offered a fairly brutal assessment of the US position, saying America "largely invented" robotics but had grown "tired and exhausted" before AI arrived to supply the missing brain.
That is the CEO of the most valuable company in the world telling you the robot supply chain runs through China, and that his robot business depends on selling into it.
The squeeze that's coming
Three pressures are converging, and NVIDIA is standing in the middle of all of them.
From Washington. The July 29 FCC ban on imports of advanced foreign-made robots showed the US is now willing to treat robots themselves as a security category, not just the chips inside them. If physical AI becomes load-bearing infrastructure for Chinese robotics, the logic that produced training-chip controls applies almost verbatim to GR00T and Jetson. There is no principled distinction between "compute that trains a model in a data center" and "compute that runs a model inside a machine with cameras walking around a factory" — only a regulatory lag.
From Beijing. China is building domestic alternatives across chips, memory, and AI systems with considerable urgency. NVIDIA's China robotics revenue is best understood as a window, not an annuity. Every quarter that window stays open is a quarter of revenue; it is also a quarter in which Chinese customers train their engineers on somebody else's stack, which cuts both ways.
From the market itself. Unitree's IPO filing on Shanghai's STAR Market — seeking roughly $609 million — signals that Chinese humanoid makers are approaching the capital markets phase, with the volume and margins to fund their own silicon roadmaps eventually.
What this means for the humanoid sector
Three takeaways for anyone tracking this industry.
- Follow the infrastructure revenue, not the robot demos. Humanoid unit shipments remain tiny. Physical AI infrastructure spend is already a $10 billion market. The money is being made one layer below where the attention is.
- The brain layer is consolidating faster than the body layer. Dozens of companies are building humanoids. A handful are building the compute and models they run on, and one of those has a decisive lead. That asymmetry will shape margins across the whole sector ... humanoid makers may find themselves in the position PC OEMs occupied relative to Intel.
- Policy is now a primary variable in robotics valuations. A single export-control rule change could remove a large slice of NVIDIA's physical AI growth and simultaneously accelerate a Chinese domestic stack. For humanoid companies choosing a compute platform today, that's a supply-chain decision with a two-to-three year shadow.
Written by
I tell stories with data and build startups. I've raised angel and VC funds, led software companies and startups, and chronicled the rise of the mobile economy. Currently I write for Forbes, host several podcasts, and am starting up HumanoidMegaHub. I was TUNE's mobile economist and Singular's VP of Insights, built the research team at VentureBeat, and managed teams creating software for partners like Intel and Disney. In past lives I built websites, e-learning platforms, and mobile apps, and, a long time ago, was named to Folio's top 100 of the media industry's "most innovative entrepreneurs and market shaker-uppers.”